Abstract:It is known that any continuous multivariate function can be represented exactly by a composition functions of a single variable - the so-called Kolmogorov-Arnold representation. It can be a convenient tool for tasks where it is required to obtain a predictive model that maps some vector input of a black box system into a scalar output. In this case, the representation may not be exact, and it is more correct to refer to such structure as the Kolmogorov-Arnold model (or, as more recently popularised, 'network'). Construction of such model based on the recorded input-output data is a challenging task. In the present paper, it is suggested to decompose the underlying functions of the representation into continuous basis functions and parameters. It is then proposed to find the parameters using the Newton-Kaczmarz method for solving systems of non-linear equations. The algorithm is then modified to support the parallelisation. The paper demonstrates that such approach is also an excellent tool for data-driven solution of partial differential equations. Numerical examples show that for the considered model, the Newton-Kaczmarz method for parameter estimation is efficient and more robust with respect to the section of the initial guess than the straightforward application of the Gauss-Newton method. Furthermore, numerical experiments show that the proposed approach performs faster than neural network training to the same accuracy.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to construct an effective Kolmogorov - Arnold model for identifying parameters from input - output data and to demonstrate the application of this model in solving nonlinear partial differential equations. Specifically: 1. **Propose a more general function decomposition method**: - The paper proposes a more general decomposition method for the basis functions in the Kolmogorov - Arnold representation through continuous basis functions and parameters. This enables the model to better adapt to different application scenarios. 2. **Use the Newton - Kaczmarz method to efficiently solve for parameters**: - The paper suggests using the Newton - Kaczmarz method to solve the nonlinear equations in the system in order to find the model parameters. This method has higher efficiency and robustness compared to the traditional Gauss - Newton method. 3. **Demonstrate the application of the Kolmogorov - Arnold model in solving nonlinear partial differential equations**: - The paper shows that the Kolmogorov - Arnold model, as a data - driven method, can effectively solve nonlinear partial differential equations. Numerical experiments indicate that this method is comparable to neural networks in terms of accuracy, but has more advantages in terms of computational time and resource requirements. ### Specific problems and solutions 1. **Generality of function decomposition**: - **Problem**: How to perform a more general decomposition of the basis functions in the Kolmogorov - Arnold representation? - **Solution**: The paper proposes a decomposition method based on continuous basis functions and parameters, making the model more flexible and able to adapt to different application scenarios. 2. **Efficiency of parameter estimation**: - **Problem**: How to efficiently identify model parameters from input - output data? - **Solution**: The paper introduces the Newton - Kaczmarz method to solve the nonlinear equations, thereby efficiently finding the model parameters. This method is more robust in the selection of initial guesses and is more efficient than the traditional Gauss - Newton method. 3. **Application in solving nonlinear partial differential equations**: - **Problem**: How to use the Kolmogorov - Arnold model to solve nonlinear partial differential equations? - **Solution**: The paper shows how to apply the Kolmogorov - Arnold model to data - driven nonlinear partial differential equation solving. Through numerical experiments, it is proved that this method is comparable to neural networks in terms of accuracy, but has more advantages in terms of computational time and resource requirements. ### Mathematical formulas - **Kolmogorov - Arnold representation**: \[ F(X_1, X_2, \ldots, X_m)=\sum_{k = 1}^{2m+1}\Phi_k\left(\sum_{j = 1}^m f_{kj}(X_j)\right) \] - **Parameter update formulas**: \[ H_{kjp}^{q + 1}=H_{kjp}^q-\mu\frac{L}{\zeta}\frac{\partial L}{\partial H_{kjp}} \] \[ G_{kl}^{q + 1}=G_{kl}^q-\mu\frac{L}{\zeta}\frac{\partial L}{\partial G_{kl}} \] where \(\mu\) is the regularization parameter and \(\zeta\) is the squared norm of the gradient: \[ \zeta=\sum_{k = 1}^{2m+1}\sum_{j = 1}^m\sum_{p = 1}^n\left(\frac{\partial L}{\partial H_{kjp}}\right)^2+\sum_{k = 1}^{2m+1}\sum_{l = 1}^s\left(\frac{\partial L}{\partial G_{kl}}\right)^2 \]

Construction of the Kolmogorov-Arnold representation using the Newton-Kaczmarz method

The Proof of Kolmogorov-Arnold May Illuminate Neural Network Learning

Chebyshev Polynomial-Based Kolmogorov-Arnold Networks: An Efficient Architecture for Nonlinear Function Approximation

The Kolmogorov-Arnold representation theorem revisited

A Survey on Kolmogorov-Arnold Network

Sinc Kolmogorov-Arnold Network and Its Applications on Physics-informed Neural Networks

Anamnesic Neural Differential Equations with Orthogonal Polynomial Projections

Kolmogorov-Arnold Networks for Time Series: Bridging Predictive Power and Interpretability

Kolmogorov--Arnold networks in molecular dynamics

KAN: Kolmogorov-Arnold Networks

Exploring the Potential of Polynomial Basis Functions in Kolmogorov-Arnold Networks: A Comparative Study of Different Groups of Polynomials

KAT to KANs: A Review of Kolmogorov-Arnold Networks and the Neural Leap Forward

Kolmogorov Arnold Informed neural network: A physics-informed deep learning framework for solving PDEs based on Kolmogorov Arnold Networks

NEW DATA ANALYSIS ALGORITHMS BASED ON SPLINE VERSIONS OF KOLMOGOROV ARNOLD NETWORKS

Kolmogorov Arnold Informed neural network: A physics-informed deep learning framework for solving forward and inverse problems based on Kolmogorov Arnold Networks

A Bregman-Kaczmarz method for nonlinear systems of equations

Data-driven model discovery with Kolmogorov-Arnold networks

Generalization Bounds and Model Complexity for Kolmogorov-Arnold Networks

Deep Learning Alternatives of the Kolmogorov Superposition Theorem

Graph-theoretic algorithms for Kolmogorov operators: Approximating solutions and their gradients in elliptic and parabolic problems on manifolds

Physics Informed Kolmogorov-Arnold Neural Networks for Dynamical Analysis via Efficent-KAN and WAV-KAN